CN-121977431-A - High-precision measurement method of viscoelastic flexible strain sensor based on PINN-LSTM

Abstract

The invention belongs to the technical field of strain sensors, and relates to a high-precision measurement method of a viscoelastic flexible strain sensor based on PINN-LSTM, which comprises the steps of firstly, carrying out tensile test on the flexible strain sensor to obtain a dataset containing time, strain, stress and resistance; the method comprises the steps of preprocessing a data set, preprocessing the data set, namely smoothing denoising, feature enhancement, resampling and normalization, reconstructing a TCN-Attention-LSTM hybrid model, wherein the TCN-Attention-LSTM hybrid model comprises a time sequence convolution layer, an Attention mechanism layer, a physical decoder and a long-period memory neural network layer, training the hybrid model by the preprocessed data set, and finally estimating the strain measurement value of the flexible strain sensor with high precision by the trained hybrid model. The invention can accurately capture the hysteresis nonlinearity and repeatability response rule of the sensor under the cyclic load condition, and realizes the average estimation accuracy of more than 95%.

Inventors

• WANG XI
• SHI YAYU
• PAN YANG
• LI QIAO
• SHEN BO

Assignees

• 东华大学

Dates

Publication Date

20260505

Application Date

20260120

Claims (9)

1. 1. A high-precision measurement method of a viscoelastic flexible strain sensor based on PINN-LSTM is characterized by comprising the following steps: (1) Performing a tensile test on the flexible strain sensor to obtain a data set containing time, strain, stress and resistance; (2) Preprocessing a data set; The preprocessing comprises smooth denoising, characteristic enhancement, resampling and normalization, wherein the characteristic enhancement refers to the construction of a series of derivative characteristics based on the coupling relation between a mechanical signal and an electrical signal, and the derivative characteristics comprise a time-displacement slope, an electrical signal-displacement slope, a force-displacement slope and a viscoelasticity characteristic; (3) Constructing a TCN-Attention-LSTM hybrid model; The TCN-Attention-LSTM hybrid model comprises a time sequence convolution layer, an Attention mechanism layer, a physical decoder and a long-term memory neural network layer, wherein the input of the time sequence convolution layer is a preprocessed data set, the output of the time sequence convolution layer is divided into two paths and simultaneously transmitted into the Attention mechanism layer and the physical decoder, the output of the Attention mechanism layer is transmitted into the long-term memory neural network layer, and the output of the physical decoder is six core physical parameters constrained in a certain range: 、 、 、 、 、 wherein 、 、 、 Is the stiffness coefficient of the spring, 、 Calculating to obtain a physical equation residual error loss according to the six core physical parameters, and optimizing the weight of the long-term and short-term memory neural network layer through gradient feedback together with the data fitting loss; (4) Training a TCN-attribute-LSTM hybrid model by utilizing the preprocessed data set; (5) And (3) performing high-precision estimation on the strain measurement value of the flexible strain sensor by using the trained TCN-Attention-LSTM hybrid model.
2. 2. The high-precision measurement method of the viscoelastic flexible strain sensor based on PINN-LSTM according to claim 1, wherein the flexible strain sensor is subjected to tensile test by using a universal material tester in the step (1).
3. 3. The high-precision measurement method of the viscoelastic flexible strain sensor based on PINN-LSTM as claimed in claim 1, wherein in the step (2), smooth denoising is performed by adopting a moving average method, resampling is performed by adopting a linear interpolation method, and normalization means that features are scaled to a [0,1] interval by adopting a Min-Max normalization method.
4. 4. The high-precision measuring method of PINN-LSTM based viscoelastic flexible strain sensor as set forth in claim 1, wherein in step (3) 、 、 、 、 、 The following relationships are satisfied: ; ; Wherein, the In order to be strained the material is, In the event of a stress being applied to the substrate, Representing the resistance.
5. 5. The method for high-precision measurement of a viscoelastic flexible strain sensor based on PINN-LSTM as set forth in claim 4, wherein the method is based on 、 、 、 、 、 Calculating to obtain the residual error loss of the physical equation The formula is: ; ; 。
6. 6. The high-precision measurement method of a viscoelastic flexible strain sensor based on PINN-LSTM according to claim 5, wherein the residual loss of the physical equation and the data fitting loss form a total loss function by weighted summation, specifically as follows: ; Wherein, the Indicating the total loss of the total of the components, The data fit loss is represented by a data fit loss, Representing the physical equation residual loss.
7. 7. The high-precision measurement method of a viscoelastic flexible strain sensor based on PINN-LSTM according to claim 1, wherein in step (4) model training is performed by Adam optimizer, batch size is set to 64, training round is 200, and early stop strategy is introduced to prevent overfitting.
8. 8. The high-precision measurement method of the viscoelastic flexible strain sensor based on PINN-LSTM as set forth in claim 7, wherein the training process adopts a two-stage strategy, wherein the optimal weight of the physical decoder is obtained in the first stage of training, and the critical super-parameters of the model are automatically optimized by adopting a Bayesian algorithm in the second stage of training, and then the model architecture is optimized by adopting a data-physical mixing loss function formed by residual loss of a physical equation and data fitting loss.
9. 9. The high-precision measurement method of the viscoelastic flexible strain sensor based on PINN-LSTM as set forth in claim 8, wherein when the Bayesian algorithm is adopted to automatically optimize the key super parameters of the model, the defined super parameter search space comprises learning rate, taking the values of [1e-4, 1e-2], LSTM unit number of [16,512], discarding rate of [0.1, 0.5].

Description

High-precision measurement method of viscoelastic flexible strain sensor based on PINN-LSTM Technical Field The invention belongs to the technical field of flexible strain sensors, and relates to a high-precision measurement method of a viscoelastic flexible strain sensor based on PINN-LSTM. Background The flexible strain sensor has the core advantages of excellent stretchability, light weight, high sensitivity and the like, has wide application prospects in the fields of medical health, intelligent robots, wearable equipment and the like, and is particularly embodied in a plurality of key directions such as motion capture, state monitoring, man-machine interaction and the like. Viscoelastic materials are key components of flexible strain sensors, whose constitutive behavior directly affects the measurement performance of the sensor. The inherent mechanical relaxation characteristic of the material can cause obvious stress relaxation and creep phenomena of the sensor in the use process, so that obvious signal hysteresis and nonlinear response are introduced. Furthermore, the repeatability of the sensor is also negatively affected due to the energy dissipation effect. Therefore, although the viscoelastic material endows the sensor with good flexibility and stretching capability, measurement error sources such as large hysteresis, poor repeatability, obvious nonlinearity and the like are brought at the same time, so that the application of the viscoelastic material in a scene requiring high-precision dynamic monitoring is limited to a certain extent. In order to improve the measurement accuracy of the flexible strain sensor, building a viscoelastic constitutive model capable of accurately representing the mechanical behavior of a core material of the flexible strain sensor has become one of the key scientific problems in the field. Early researches were mainly based on classical viscoelastic theory, and the basic construction method is to combine an elastic element (spring) and a viscous element (viscous pot) in series or in parallel to form basic models such as a Maxwell model （Maxwell, J. C. (1867). On the dynamical theory of gases. Philosophical Transactions of the Royal Society of London, 157, 49–88.）, a Kelvin-Voigt model （Voigt, W. (1890). Über das Verhältnis zwischen den beiden Elastizitätskonstanten isotroper Körper. Annalen der Physik, 274(12), 573–587.） and the like. Further, by combining the two models in different ways, more complex viscoelastic models such as the Zener model (Zener, c. (1948) & ELASTICITY AND ANELASTICITY of metals & University of Chicago press.) and the Burgers model （Burgers, J. M. (1935). Mechanical considerations—Model systems—Phenomenological theories of relaxation and of viscosity. In First Report on Viscosity and Plasticity (pp. 5–67). Nordemann Publishing Co.） were developed. To accurately describe material behavior over a wider time or frequency range in the middle and later of the 20 th century, researchers introduced multiple Maxwell branches connected in parallel to construct a generalized Maxwell model (Ferry, j.d. (1980). Viscoelastic Properties of Polymers (3 rd ed.). Wiley.) and fitted experimental data using Prony series to determine each branch parameter （Tschoegl, N. W. (1989). The Phenomenological Theory of Linear Viscoelastic Behavior. Springer-Verlag.）. on the other hand, as the requirements for modeling material time correlation, memory effects, and long-time scale behavior increase, traditional integer-order combination models have difficulty in compromising between accuracy and parameter simplicity. Therefore, fractional derivative theory （Bagley, R. L., & Torvik, P. J. (1983). A theoretical basis for the application of fractional calculus to viscoelasticity. Journal of Rheology, 27(3), 201–210.） is introduced as popularization of the traditional damping element, and models （Mainardi, F. (2010). Fractional Calculus and Waves in Linear Viscoelasticity: An Introduction to Mathematical Models. Imperial College Press.）. of Maxwell, kelvin-Voigt, zener, burgers and the like of fractional versions are derived, so that the capability of the viscoelastic theory in the aspects of physical mechanism explanation and mathematical expression accuracy is remarkably improved. However, such model parameter identification is highly dependent on specific experimental data, and generalization ability under high strain rate or large deformation conditions remains to be improved. The electronic fabric sensor prepared from the conductive composite material has inherent defects, and the electric signal of the electronic fabric sensor shows obvious hysteresis and drift phenomena along with time, so that the sensor continuously has instantaneous errors, and the accuracy requirement of real-time measurement is difficult to meet. Thus, research into resistance relaxation behavior has become one of the keys to improve the performance of flexible sensors. Literature 1（Phenomenologica